Running a Recipient List Email Experiment with marginal
A recipient list email experiment splits your audience and sends competing variants to measure what actually moves opens and clicks. marginal lets an AI agent run that whole loop — from variant generation to launch to lift analysis — through a single hosted email MCP server.
What a recipient list email experiment is
At its core, a recipient list email experiment takes one audience and divides it across two or more message variants. Each slice of the list gets a different subject line (or send), and you compare engagement to decide which performs best. Done well, it removes guesswork: you replace opinions about what sounds good with measured open and click rates from real recipients.
marginal handles the experiment plumbing so you don't wire up cohort assignment, send orchestration, or tracking yourself. You supply the list and the idea; the platform manages the split and reports within-test lift between variants.
- Define one audience, multiple competing variants
- Randomized assignment across the recipient list
- Open and click tracking captured per variant
- Within-test lift metrics to compare performance
Driving the experiment from an AI agent
Because marginal is an email MCP server, any MCP-capable assistant can run the experiment as a conversation. AI agent email marketing means you describe the campaign in plain language and the agent calls the right tools in sequence — no dashboard clicking, no manual exports.
The four MCP tools map cleanly to the experiment lifecycle, so the agent can carry a test from idea to recommendation without leaving your editor or chat.
- generate_variants — draft subject line options to test
- launch_test — split the recipient list and send
- get_results — pull open/click and lift numbers
- recommend_next — suggest the follow-up based on outcomes
Designing a clean subject line A/B test
The most common recipient list email experiment is a subject line A/B test, where the only difference between cohorts is the subject. Keeping everything else constant — body, send time, sender — isolates the subject's effect so your lift number means something.
A few practical habits keep results trustworthy. Ensure each cohort is large enough to detect a real difference, change one element at a time, and let the test collect enough opens before declaring a winner. marginal's lift metrics show the gap between variants so the agent can act with confidence.
- Test one variable per experiment for clear attribution
- Use comparable cohort sizes from the same list
- Wait for meaningful open volume before deciding
- Feed the winner into your next campaign
Getting started on marginal
marginal is hosted at https://marginal.sh/mcp — there's nothing to self-host. Connect a supported client such as Cursor, Claude Desktop, Claude Code, Windsurf, Cline, Continue, Codex, or Zed, authenticate with your Bearer API key, and start running experiments.
The free tier covers 100 experiments per month, which is enough to validate a subject line A/B test workflow before scaling. See the docs at https://marginal.sh/docs/ for tool details and examples.
- Endpoint: https://marginal.sh/mcp
- Auth via Bearer API key (marg_live_...)
- Free tier: 100 experiments/month
- Works across major MCP-capable clients
Get started with marginal
marginal is a hosted email marketing MCP server at marginal.sh. Sign up free, create an API key, and connect https://marginal.sh/mcp from Cursor, Codex, or Claude Desktop.
- 100 experiments/month on the free tier
- Four MCP tools: generate_variants, launch_test, get_results, recommend_next
- Listed in the MCP Registry as sh.marginal/mcp